Combining Breast Cancer Risk Prediction Models

Abstract

Accurate risk stratification is key to reducing cancer morbidity through targeted screening and preventative interventions. Multiple breast cancer risk prediction models are used in clinical practice, and often provide a range of different predictions for the same patient. Integrating information from different models may improve the accuracy of predictions, which would be valuable for both clinicians and patients. BRCAPRO is a widely used model that predicts breast cancer risk based on detailed family history information. A major limitation of this model is that it does not consider non-genetic risk factors. To address this limitation, we expand BRCAPRO by combining it with another popular existing model, BCRAT (i.e., Gail), which uses a largely complementary set of risk factors, most of them non-genetic. We consider two approaches for combining BRCAPRO and BCRAT: (1) modifying the penetrance (age-specific probability of developing cancer given genotype) functions in BRCAPRO using relative hazard estimates from BCRAT, and (2) training an ensemble model that takes BRCAPRO and BCRAT predictions as input. Using both simulated data and data from Newton-Wellesley Hospital and the Cancer Genetics Network, we show that the combination models are able to achieve performance gains over both BRCAPRO and BCRAT. In the Cancer Genetics Network cohort, we show that the proposed BRCAPRO + BCRAT penetrance modification model performs comparably to IBIS, an existing model that combines detailed family history with non-genetic risk factors.

Keywords: BCRAT; BRCAPRO; ensemble learning; model aggregation; stacking.

Figure A1

BCRAT cause-specific hazard of breast cancer for White women in the general population (${\tilde{\lambda }}_B\left(t\right)={\tilde{\lambda }}_{B,\mathbf{0}}\left(t\right)/(1-AR\left(t\right))$) and BRCAPRO cause-specific hazard of breast cancer for White female non-carriers (${\lambda }_B^0\left(t\right)$).

Figure 1

Inputs to BRCAPRO and BCRAT and their overlap.

Figure 2

Calibration plots by decile of risk for five-year predictions in a simulated dataset with 45,557 counselees (724 cases). For each model, we grouped individuals by decile of risk and plotted the observed proportion of women who developed cancer (with 95% Wilson CI) versus the predicted probability (sum of risk predictions) within each decile.

Figure 3

Scatter plots, density plots, and pairwise correlations of five-year predictions in the CGN data. Red corresponds to data from cases, blue corresponds to data from controls (individuals who did not develop breast cancer within five years), and beige corresponds to data from counselees censored before five years. Corr: Pearson correlation, cens: counselees censored before five years, ctrl: controls. The lower diagonal panels show scatter plots of the predictions from each pair of models. For example, the panel in the second row, first column has predictions from BRCAPRO + BCRAT (E) on the x-axis and predictions from BRCAPRO + BCRAT (M) on the y-axis. The diagonal panels show density plots of the predictions from each model, stratified by case-control status. For example, the panel in the first row, first column shows the distribution of predictions from BRCAPRO + BCRAT (M). The upper diagonal panels show the Pearson correlations between predictions from each pair of models. For example, the first row, second column shows the overall correlation, correlation among cases, correlation among censored counselees, and correlation among controls for predictions from BRCAPRO + BCRAT (M) and BRCAPRO + BCRAT (E). *** = p < 0.001.

Figure 4

Calibration plots by decile of risk for five-year predictions in CGN. For each model, we grouped individuals by decile of risk and plotted the observed proportion of women who developed cancer (with 95% Wilson CI) versus predicted probability (sum of risk predictions) within each decile. In computing the observed proportions, the inverse probabilities of the censoring weights were used to account for censoring.